Scientists say AI is falling for 'alien hoaxes' too easily — and that's a problem for research

• Researchers have spotted problems with AI pattern matching in science data
• It could mean false flags for signatures of life on other planets
• AI can still be useful, but checks need to be built in

One of the ways AI can be most helpful is in trawling through masses of scientific data that human researchers don't have time to analyze, looking for patterns — but this use-case is now proving problematic when it comes to the search for life beyond our planet.

A new study from researchers at Michigan State University suggests that AI systems can be too easily fooled into identifying signatures of life out in the universe where none exist. We need these flags to be accurate to know where to point our telescopes next, so it's important that the detection processes work.

The researchers set up a digital simulation including a key sign of life: the ability for molecules to replicate and mutate. Software was used to generate tens of thousands of digital organisms with and without this ability, which where then used to trail a neural network to spot the difference with an accuracy rating of 99.7%.

When the neural network was pointed towards data it hadn't previously seen, however, the AI's life-spotting skills fell apart. The researchers started with a digital organism that couldn't copy itself, which the AI correctly identified, then began making small edits and asking the AI to check again.

Essentially, as the AI was nudged out of its comfort zone of training data, it started seeing life where there wasn't any. "No matter what sequence of commands we started with, we were able to fool the AI 100% of the time," said Ankit Gupta, one of the researchers.

Space and beyond

It's worth bearing in mind the limitations of this research: these tests were carried out in an artificial, digital simulation, and so didn't rely on any real data. The researchers were deliberately searching for errors too, rather than letting them happen by chance.

However, the study methods are solid enough to be concerning. The worry is that a Mars rover or a deep-space telescope could identify a life signature with a high degree of confidence, without necessarily having a human in the loop to check.

The researchers found there were a vast number of sequences that could trip up the AI too, meaning the risk of a mistake is more likely. While the digital organisms incorrectly identified by the neural network were close to what it had been trained to spot, they weren't full matches — despite the AI thinking they were.

These issues could crop up outside of space exploration too. The same errors might appear when looking for patterns in medical scans, security camera footage, and everywhere else the technology is used.

That said, the researchers are keen to emphasize that AI can still be useful in these scenarios — it just needs careful checks and supervision. "AI has an Achilles' heel: it can see a pattern and completely misclassify it," said Christoph Adami, one of the team. "There needs to be a human in the loop."